Real-Time Classification of Invasive Plant Seeds Based on Improved YOLOv5 with Attention Mechanism

Seeds of exotic plants transported with imported goods pose a risk of alien species invasion in cross-border transportation and logistics. It is critical to develop stringent inspection and quarantine protocols with active management to control the import and export accompanied by exotic seeds. As a result, a method for promptly identifying exotic plant seeds is urgently needed. In this study, we built a database containing 3000 images of seeds of 12 invasive plants and proposed an improved YOLOv5 target detection algorithm that incorporates a channel attention mechanism. Given that certain seeds in the same family and genus are very similar in appearance and are thus difficult to differentiate, we improved the size model of the initial anchor box to converge better; moreover, we introduce three attention modules, SENet, CBAM, and ECA-Net, to enhance the extraction of globally important features while suppressing the weakening of irrelevant features, thereby effectively solving the problem of automated inspection of similar species. Experiments on an invasive alien plant seed data set reveal that the improved network model fused with ECA-Net requires only a small increase in parameters when compared to the original YOLOv5 network model and achieved greater classification and detection accuracy without affecting detection speed.